Changing News Media Landscape in South Korea

Hongjun Lim∗ Choongho Chung∗ Jihee Kim∗∗Juho Kim† Sue Moon† Meeyoung Cha∗

∗Graduate School of Culture Technology, KAIST, South Korea∗∗College of Business, KAIST, South Korea†School of Computing, KAIST, South Korea

{limhj159, thegenuine, jiheekim, juhokim, sbmoon, meeyoungcha}@kaist.ac.kr

ABSTRACTA recent disclosure of a presidential scandal and the follow-ing development of subsequent events have been a majornews topic in South Korea. We conducted a data-drivenstudy to examine public reactions to the scandal and theireffect on the Korean media landscape. Our analysis is basedon 59,224 news articles published by five popular newsroomsthat received a total of 47,906,770 Likes on Facebook. Thedata reveal notable changes in media ranks throughout thescandal, where a relatively young TV news network out-grew its audience over other half-a-century old establishednewspapers in less than a month. Topical similarity of newsheadlines also remained high over an extended period oftime, despite the varying political stance of each newsroom.The topical similarity further shows a gradual divergence,indicating re-positioning of media stance takes place overtime. Implications of these findings and suggestions for fu-ture directions are discussed based on data analysis of thisextraordinary event.

1. INTRODUCTIONWith growing digital news consumers and a plethora of

platforms through which news stories disseminate, the globalnews media landscape has been changing fast. Accordingto studies, digital and social media platforms like YahooNews, Facebook, and YouTube have democratized mediaand, as a result, digital-only platforms like Buzzfeed nowhave an equal chance to spread stories as any establishedplatforms [9]. Changes in the news media landscape aresometimes driven by innovations in newsrooms themselves,for instance, the uptake of the Huffington Post via a blog-based strategy and the New York Times’ shifts to visual-centric and socially share-worthy stories.

Another source of changes in media landscape is large so-cietal events. These events incur significant changes in pat-terns of news cycles, in the way the public with prior knowl-edge of the events reacts to the follow-up news [6], andin dynamics between media outlets. Elections and political

c©2017 International World Wide Web Conference Committee(IW3C2), published under Creative Commons CC BY 4.0 License.WWW’17 Companion, April 3–7, 2017, Perth, Australia.ACM 978-1-4503-4914-7/17/04.http://dx.doi.org/10.1145/3041021.3054762

.

scandals affect the 24-hour news cycle and increase audi-ence interactions. Moreover, there had been cases where anews outlet rose to a position of power by reporting newsscoops for a major event and providing insights in follow-up reports, ending up with a wide public trust and favor.The role of the Washington Post on Watergate (1972), andof Al Jazeera on Arab Spring (2010) provided opportunitiesfor the two media to be recognized by the public and gainauthority, either domestically or internationally [7, 10].

This paper reports one such transformation point in theKorean media landscape that started over a political scandalin 2016 and is still on-going as of the writing of this paper inFebruary 2017. The scandal, involving allegations for the ex-tortion, abuse of power, and bribery of the Korean president,has led to the largest ever protest in the country on Decem-ber 3rd, with the estimated number of participants over 2.3million (or 4.5% of South Korea’s population) gathering forpeaceful demonstrations. The cumulative count of partici-pants from late October to December adds up to more than10 million. A motion of impeachment has been approvedon December 9th in the National Assembly, and awaits finalruling by the Constitutional Court.

Coverage of this scandal by Korean media is noteworthyfor several reasons. First, the event increased news audi-ence. Television viewer ratings on news programs went upby 1.5 times and the total reactions on news stories on socialmedia went up by 1.9 times. Update of news is particularlyinteresting given the trend that people’s trust in mass me-dia has been declining in Korea. Second, news media rankschanged. The Korean news influence until recently has beendominated by established newsrooms such as Chosun Daily,Joongang Daily, Kyunghyang, and MBC News that werefounded half a century or little over a century ago. Thepolitical scandal, however, changed this ranking. A cablenetwork-based newsroom, JTBC, that was only establishedseveral years took the lead over other half-a-century old es-tablished newsrooms. JTBC was the first to disclose a con-crete evidence that sparked the presidential scandal. Itssheer focus and dedication to provide consistent high-qualitycoverage of the scandal might have helped gain the largestaudience in terms of both television viewer ratings and socialmedia reactions.

In this paper, we collect data from the Facebook pages ofmajor newspaper publishers in South Korea in order to ex-amine the changing media landscape over this notable polit-ical scandal. To quantify the differences in topical coverageand audience reactions across media outlets, we compile aset of topics seen in news headlines for each media outlet.

1521

Then using the generated topic sets, we compare topical sim-ilarity across pairs of media outlets and visualize their topicnetworks. We also examine how each news outlet gained orlost their influence after the scandal.

This paper makes the following observations:

1. Similarity of news headline topics can be an alternativemetric to quantify the degree to which media outletsconcurrently cover a large societal event.

2. A visualization helps highlight topical similarity as wellas dissimilarity across media outlets, which could beused to alleviate the biasing effect of filter bubbles byexposing its audience to diverse media stance.

3. A data-driven method adopted in this study uniquelycaptures the unprecedented scale of topical conver-gence in Korean media regarding the political scan-dal and a resulting rise of a new winner—a relativelyyoung news network—in terms of audience influence.

This work is limited in that Facebook news audience isnot representative of all news audience in Korea. However,the trends seen on this platform help us observe prominentchanges in media ranks, which before the social media erawould have been hard to observe closely. Hence our studyis a preliminary yet a meaningful attempt to witness thechanging dynamics in the Korean media landscape. Ouranalysis on social media reactions on news stories providesinsights on how news scoops are interpreted and conveyed toonline audience differently, which are critical to uncoveringhow people read and react to news online.

2. DATA METHODOLOGY

2.1 Data CollectionNews posts studied in this paper were accessed through

the Facebook Graph API over a single day period on Jan-uary 10th, 2017. We gathered a total of 166,131 posts, whichare the entire set of public posts of five news media outlets inSouth Korea. They include two of the popular right-leaningmedia, Chosun Daily and Joongang Daily, and two of thepopular left-leaning media, Hankyoreh and Kyunghyang, re-spectively. Also included in this list is a relatively newlyemerged cable news network, JTBC, that has become no-table for releasing a number of exclusive news and scoopson the event.

Media Data since Posts Total LikesJTBC 2011/11/28∼ 17,323 12,612,111

Chosun Daily 2010/04/16∼ 26,032 25,210,818Joongang Daily 2010/11/09∼ 47,089 10,634,283Kyunghyang 2012/05/28∼ 33,001 31,655,183Hankyoreh 2010/08/05∼ 42,686 27,194,615

Total 2010/04/16∼ 166,131 107,307,010

Table 1: Summary of the dataset

Table 1 displays key information about the gathered dataindicating the data period, post count, and the aggregateLikes count of news posts. The five outlets opened theirFacebook pages at different points in time, attributing todifferent time range of data between the outlets. For eachnews post, we collect information about the timestamp (i.e.,

time when a post was uploaded to Facebook), main text(i.e., short text description about any linked content), newsheadline (i.e., text headline of the news article that is linkedon the Facebook post, if any), news video (i.e., video contentthat is directly uploaded to the page, if any), news image(i.e., image content that is directly uploaded, if any), aswell as the count of Likes and Comments. To ensure thatall posts had enough time to circulate within Facebook, weanalyze only those news stories posted in 2016.

Three kinds of analysis methods are utilized in this studyOne is the Likes count for each Facebook post. Another istopic similarity of news posts. When examining news topicsof each media outlet, we do not restrain by the aforemen-tioned circulation time and use all recent data. Last is thetopic network visualization, which assists understanding ofthe varying keywords covered by each media outlet. Weexplain these methods in detail.

2.2 Topic AnalysisTo analyze topics, we combined the text data from the

main description and the linked news headline, if any ex-isted, for each post. A majority of posts (78.9%) containedat least one news link with a headline; the remaining postswere featuring news videos or images and did not containany external links to news sources, in which case only themain text description was used for topic analysis. Next, thecombined text was treated with the Korean morphologicalanalyzer Komoran (via an external Java package) to tok-enize words and remove stop words (e.g., ‘a’, ‘the’)1. Sincethe gathered posts were produced by newsrooms, languageusage was clean and most content could be tokenized prop-erly. To better assist the tokenizing process, we have manu-ally added a set of keywords related to the event for namedentity recognition based on the user dictionary provided byKomoran. For instance, we added keywords indicating thename of people and organizations involved in the scandal tothe user dictionary such as ‘Choi Soon-sil’ and ‘Federationof Korean Industries’.

After tokenizing, we are left with a potential set of key-words in the form of nouns, verbs, and adjectives in rootforms. We assume a part of the text description and newsheadlines published by media outlets can be a candidate fora topic, rather than to extract subtext or the hidden mean-ing from the text. As one such approach, we further utilizednouns as possible keywords in this research [4]. In Korean,nouns are classified into several types: common nouns (e.g.,‘cookie’, ‘bakery’), proper nouns (e.g., ‘Oreo’, ‘Chips Ahoy’),dependent nouns (e.g., ‘what’), and rhetoric and pronouns(e.g., ‘he’, ‘she’). Among these noun types, we further con-sider proper nouns and common nouns as candidate topicsets, as they are the only nouns that can have independentmeanings. Each post, hence, may contain one or more nounsin its topic set. On average Facebook posts contained 10.46nouns as topics.

We then utilized TF-IDF (Term Frequency-Inverse Doc-ument Frequency) on the pre-processed topic sets (nouns)to identify prominent topics. TF-IDF is a measure of howimportant a given word is in a document and is normallyused to calculate weighted similarity between a documentand a term in a collection. This method was applied to dif-ferentiate topical differences across media outlets. We calcu-

1Komoran 2.0 Korean morphological analyzerhttp://www.shineware.co.kr/?page_id=835

1522

Figure 1: Timeline of the 2016 South Korea’s presidential scandal seen by the daily total Likes on Facebooknews posts. The Likes count has been aggregated across all five news outlets.

lated similarity of the identified topic sets by constructing aterm vector for each media page via a cosine function. Thetopical similarity between two news media outlets hence isdetermined by a value ranged between 0 and 1, where thevalue of 1 indicates a complete overlap in the topics coveredand the value of 0 means otherwise.

2.3 Topic VisualizationVisualization was used to illustrate differences in topical

trends across media outlets. Here, rather than individu-ally treating each noun separately as a topic, we tried toidentify any phrases or co-occurring patterns among topicalnouns. This allows us to employ a network-based visual-ization, where a node represents a topic and an edge rep-resents a link between two topics if they are associated. Anumber of association rules can be applied here, where themain idea is to highlight closely related nouns in the topicsets. This study relied on one such heuristic based on theApriori algorithm [1], whose association rule tries to infercausal relationship. Keyword co-occurrences were examinedat the post level. The algorithm allows a single keyword ora noun in a post to appear across multiple associations. Forinstance, a keyword ‘president’ can appear as a singletonand at the same time be associated with another keywordto form a word pair (‘president’,‘impeachment’) or (‘presi-dent’,‘respond’). Such subtle difference in co-appearance ofkeywords in news headlines reveal the careful choice of wordsand political stance across media outlets.

In applying the association rule, we only considered top150 nouns from each media outlets, so that only prominentassociations are learned. We applied the arules package inR programming language [5]. The Apriori algorithm wasrun for each news post to form a topic network. Then, theresulting topic associations were visualized as a directed net-work by the Gephi tool, an open-source network analysis andvisualization software [2].

3. RESULTSEffect of the political scandal on the Korean news media

landscape is highlighted from three perspectives: popularityof news stories, shifts in media ranks, and topical similarityof news headlines over time. We also examine an increasein the production of ‘exclusive’ stories.

• The public reception of the scandal measured bycontent popularity. We first examine the temporal evo-

lution of the 2016 presidential scandal from the public re-actions and the media coverage, in turn. First, the publicreaction can be represented by the aggregate Likes countreceived on news posts. This trend is shown in Figure 1,where the news story (‘Disclosure of evidence on politicalscandal’ in the figure) published on October 24th eveningby the JTBC newsroom marks the beginning of the hugepublic interest. That news story was the very first instancefor any news media outlet to disclose a tangible piece ofevidence supporting the prior allegations on the president.Compared to the week before the scandal, news articles onFacebook received on average 1.9 times more Likes per post.While not shown in the figure, the total number of newsposts per day, however, is similar to early October in thatnone of the five media outlets produced significantly morenews posts on their Facebook page upon the event.

The public reaction to news posts shows peaks on certaindays. The largest peak appears on November 12th, when thefirst million-people protest was organized at the Gwangh-wamun Square in Seoul, which demanded the ouster of thepresident. Protests were organized regularly and voluntarilyevery Saturday and the graph shows subsequent peaks cor-responding to such events, albeit smaller than the peak onthe first million-people protest. It is also clear that some keynews events do not correlate with peaks in Figure 1. How-ever, it is still reasonable to assume that high Likes countlevels and peaks generally represent public attention to im-portant events.

We then extend our analysis to the news media’s perspec-tive on the scandal. In occurrence of a large event, we expectvarious newsrooms to report more on the event, regardless oftheir political stance. This will lead to a temporary increasein similarity in topics across news media outlets. Figure 2shows similarity of news topics between two news outlets ofopposite leaning, plotted from the beginning of 2016 to early2017. Topical similarity at each data point (appearing ev-ery 3 days) is aggregated over an overlapping 2-week movingwindow. It is easy to see that similarity in news headlinespeaks when notable societal events occur. Decreasing trendin similarity after the peak implies that newsrooms simplyshifted their focus to other new topics or are providing followup reports with different wordings due to their own inter-pretation of the major event.

Nonetheless, the political scandal in 2016 stands outamong previous peaking events in two ways. First, thescandal is associated with the highest degree of similarity.

1523

Figure 2: A one year-long trend on topical similarity of news headlines between two prominent news outlets:Chosun Daily (right-leaning media) and Kyunghyang Shinmun (left-leaning media)

Retention of high similarity levels is also unprecedented andforms a plateau rather than peaks as in other social events.Such observations indicate that effects of the scandal onmedia attention is longer lasting. This showcases a need fora deeper analysis on how news outlets adapt to the scandalover extended periods of time, which we aim in this study.

• Shifts in media ranking and effects of exclusivenews post. While the previously introduced similaritygraph calls for further analysis, the significant increase inLikes count in Figure 1 also raises question on which mediaoutlet attributes to such increase in public attention. Fig-ure 3(a) shows the average Likes count per post for eachmedia outlet along with error bars measured in standarderrors before and after the event. Like counts were calcu-lated over a non-overlapping month period before and af-ter October 24th. The political leaning of each news outletis denoted under its name as R (right-leaning) or L (left-leaning). Most platforms show an increment in audienceengagement. However, the most popular right-wing outletthat has the largest offline newspaper circulation in SouthKorea, Chosun Daily, shows a slight decrease in the Likescount. JTBC, the youngest newsroom, in contrast, morethan triples the Likes count over the same time period.

This finding demonstrates that circumstances surround-ing the political scandal did lead to a change in the SouthKorean media landscape in terms of which media the pub-lic favors. A further analysis on nature of news posts wasperformed to identify the underlying causes of this change.Among many factors that might have affected to this dy-namics, here we focus on two such causes. One is that wesaw a frequent mention of ‘exclusive’ in news headlines ofJTBC posts. Hence we assume that exclusive news postson the scandal gathers more Likes than non-exclusive news,and that any media outlet’s superiority on producing suchnews results in gathering more Likes than the other newsproviders.

Figure 3(b) shows the percentage of news that we deem as‘exclusive’ stories for each news media outlet. For this anal-ysis, Facebook posts between the time of Oct 24th to Nov24th were classified as either i) exclusive news or ii) gen-eral news. This was done by looking for explicit expressionsin the news headlines; all posts containing either ‘exclusive

news’ or ‘special news’ were classified as an the exclusivenews type, while all other posts were dubbed as the generalnews type. JTBC was the leading media when it comes toreporting more news scoops, at the selected time frame. Infact, the rate of news deemed as exclusive type across mediaoutlets were proportional to the rate of likes news outletsgathered—as seen by the reverse ordering of media outletsin Figures 3(a) and (b). The results show that JTBC’s risein popularity can be partially explained by its higher per-centage of exclusive news posts to those of other news mediaoutlets.

In fact, news posts that were of the exclusive news typeon average gained 2.17 times more Likes than the generalnews type in our data, indicating that the news audienceappreciated the media’s efforts to disclose new facts aboutthe event. We also mention that, while JTBC undoubtedlycontributed largely to reporting the scandal, other newsoutlets also covered non-negligible fraction of exclusivenews in that this event was covered by multitudes of mediaoutlets in a collaborative fashion. This could be confirmedby the varying topics of exclusive news covered by eachmedia outlet; the newly published news stories were onslightly different topics from one media outlet to another,rather than repeating of the same narrative.

• Convergence of news topic similarity. After see-ing how JTBC gained popularity, we are left with anotherquestion of how the other news media outlets, with varyingpolitical stances, reacted to the scandal. As a first step inanswering the question, we explored the change of topicalsimilarities for each news provider after the scandal. Fig-ure 4 shows two such comparisons for the two right-wingmedia (Chosun Daily and Joongang Daily) and two right-left media pairs (Chosun Daily and Kyunghyang). Topicalsimilarities, measured by the cosine similarity based on TF-IDF, news headlines between newsrooms of similar politicalleaning already exhibit high similarity before the event (i.e.,nearly 0.5 or higher), whereas the cross-pairs cover differenttopics prior to the event. Upon the scandal, all media pairsconverge in their topical similarity due to the significance ofthe event.

While omitted due to space limitation, other media pairsalso showed changes; headline topics between all media

1524

(a) Average Likes count per media (b) Percent of exclusive news per media

Figure 3: Likes count change and exclusive news percentage for each media outlet

Figure 4: Topical similarity changes for different me-dia pairs

pairs showed increased similarity to a different degree thatreflects political leanings. This finding implies that eachnews provider, originally choosing topics that fit moreamicably with its set political stance, opted to report onsimilar news topics after the political scandal. Thereby, themagnitude and gravity of the political scandal reached acertain threshold where a convergence of topics of the newsmedia with different political viewpoints took place.

• Divergence of topics over time. Having recognized asudden uptake of JTBC and the lasting impact of the scan-dal on the news media, we compare further how this newlyemerged platform compares to other newsrooms during thecourse of time after the convergence of topics among all se-lected news media. Figure 5(a) shows a change of topicalsimilarity between JTBC and the most established media,Chosun Daily. The week 0 in the x-axis represents October24th and topical similarity at each data point is calculatedin the same way as Figure 2. Topical similarity betweenthe two news providers that start as low as 0.35 rises overthe event to a peak of 0.52, at the disclosure of the politicalscandal. Basically, topical similarity increases if there ex-ists more number of same topic pairs between the two newsoutlets and such increase is moderated by TF-IDF weights,

which have differing values for different topics. After reach-ing its peak on the second week, headline topics start togradually diverge over time.

In order to look deeper into such topic divergence, we visu-alized topical networks of the two media, where Figures 5(b)and (c) are network plots of headline topics covered by JTBCand Chosun Daily between December 22th, 2016 and Jan-uary 5th,the following year. Topics that are directly relatedto the presidential scandal are marked in a bubble, whereastopics that are indirectly related or not related appear out-side the bubbles.

Figure 5(b) shows that most of the news stories publishedby JTBC are about the scandal, indicated by the large cover-age of the bubble. Prominent topics here are ‘president’, ‘im-peachment’, ‘recording’, and ‘judgment’ in the graph. Sometopics appear as not directly related to the event such as ‘Ar-ticle’, ‘Social department’. JTBC’s headlines were mostlydedicated to the scandal in the time period we examined.The visualization shown in Figure 5(c) for Chosun Daily, incontrast, shows a smaller coverage of the bubble. A numberof topics appear as not directly related to the scandal suchas ‘establishment of a new party’, ‘North Korea’, and ‘theTHAAD US army missile system’. The difference in topicalsimilarity is known to be caused by differing preferences ofeach newsroom to focus on the current biggest issue in moredetail, or to focus on other topics unrelated to the politicalscandal, respectively. [3]

4. DISCUSSIONNews content has been traditionally disseminated via plat-

forms like television, newspapers, local radio, and maga-zines. Now more people are accessing news through the In-ternet and the social web. This paper utilized data that canbe gathered from one such popular platform and analyzedit to report on news media coverage and public reactionsof a large political scandal in South Korea in 2016. In thefollowing we discuss some implications of our study and sug-gestions for future research.

First, our findings have implications to media outlets.Similarity and dissimilarity in headline topics can be utilizedin positioning the newsroom’s stance better and further beused to predicting how competing media outlets will covera given story. Newsrooms can then re-evaluate what kinds

1525

(a) News headline similarity (b) Topic network of JTBC (c) Topic network of Chosun Daily

Figure 5: Diverging topical similarity in news headlines between two news outlets: JTBC and Chosun Daily

of headlines are reacted popularly by the Facebook audi-ence based on the topic network analysis (e.g., what kindsof keywords are used in such headlines).

Second, our methodology and findings can be used forbuilding tools that assist news readers. An automated vi-sualization algorithm that plots a timeline summary of keyevents based on detection of peaks [8] and plateaus couldhelp news readers quickly catch up with trending news. Fur-thermore, the map of headline topics can provide useful in-formation for interested news readers to learn the politicalstance of different media outlets. A comprehensive visualiza-tion tool that integrates both functions can easily introducediverse perspectives to news readers, which can alleviate me-dia bias caused by filter bubbles.

Third, this research contributes to journalism research inthree major ways. One is uniquely capturing the convergingand diverging headline topics through a large scandal acrossthe right-wing and left-wing media. The degree of conver-gence and its lasting period are valid indicators in classifyingexceptional public events, while the gradual process of top-ical divergence upon the onset of a political scandal opensup a new research topic on discovering how different mediaemploy unique strategies to cope with their own set of au-dience over an extended period. Another contribution is inrecognizing an uptake of a young cable news outlet (JTBC)in the Korean news media landscape; a swift coverage ofexclusive news content likely contributed to this changingdynamics.

Finally, our analysis opens up multiple directions for fu-ture work: For instance, (1) Do other online (e.g., Twitter,news websites) and offline (e.g., newspaper readership) chan-nels exhibit similar patterns observed on Facebook? (2) Howdo media and the public interact with each other in such bigevents? How does the public discourse affect or get affectedby media? (3) Within topical convergence, what strategiesdo media outlets employ to further their political stance? (4)What do the patterns of bursts and peaks tell us about thenature of events themselves? For example, do smoother andlonger peaks reflect prolonged public interest and high sig-nificance of an event? We plan to investigate these questionsto better understand and support the intricate interplay be-tween a societal event, a media landscape, and the publicdiscourse.

5. REFERENCES[1] R. Agrawal, R. Srikant, et al. Fast algorithms for

mining association rules. In proc. of the VLDB, 1994.

[2] M. Bastian, S. Heymann, M. Jacomy, et al. Gephi: anopen source software for exploring and manipulatingnetworks. In proc. of the ICWSM, 2009.

[3] R. M. Entman. Framing: Toward clarification of afractured paradigm. Journal of Communication,43(4):51–58, 1993.

[4] V. Gupta and G. S. Lehal. A survey of textsummarization extractive techniques. Journal ofEmerging Technologies in Web Intelligence,2(3):258–268, 2010.

[5] M. Hahsler, C. Buchta, B. Gruen, and K. Hornik.arules: Mining Association Rules and FrequentItemsets, 2016. R package version 1.5-0.

[6] S. R. Kairam, M. R. Morris, J. Teevan, et al. Towardssupporting search over trending events with socialmedia. 2013.

[7] H. H. Khondker. Role of the new media in the arabspring. Globalizations, 8(5):675–679, 2011.

[8] A. Marcus, M. S. Bernstein, O. Badar, et al. Twitinfo:aggregating and visualizing microblogs for eventexploration. In proc. of the ACM CHI, 2011.

[9] N. Newman, R. Fletcher, D. Levy, and R. Nielsen.Reuters Institute Digital News Report 2016. Universityof Oxford, 2017.

[10] M. Schudson. Notes on scandal and the watergatelegacy. American Behavioral Scientist,47(9):1231–1238, 2004.

1526